Oil & Gas
Oil & Gas solution_
What problems does artificial intelligence solve?
Artificial intelligence enables relationships hidden in large amounts of data. In this scenario, due to the nature of the oil and gas production business, exploration and drilling are processes that can use machine learning techniques to automate important variable tasks and modifications.
AI applications in industry
The most common applications of artificial intelligence for seismic data processing, a fundamental part of operations in the industry, are:
- Reconstruction of seismic data with low sampling and erroneous traces.
- Characterization of reservoirs for the extraction of seismic attributes.
- Failure and fracture detection in pre-migrated seismic data.
- Automatic segmentation of the horizons in a seismic volume.
- Detection of salt reservoirs.
Business impact with artificial intelligence
The processing and analysis of seismic information is part of the exploration of a hydrocarbon producer company.
In this vertical, AI projects for data analytics are aimed at reducing time in tasks that are manual for geologists and geophysicists, in addition to improving precision in identifying horizons, faults, fractures, reservoirs, lithology, among others, which are important indicators of the presence of gas or oil.
The benefits of applying any of these solutions depend on factors specific to the company and the field to be analyzed.
Suggested AI and analytics roadmap
- Guide and align all business stakeholders about the importance and benefits of applying artificial intelligence.
- Identify problems and opportunities that generate real value for the company and prioritize needs.
- Identify the sources of information and estimate the effort to extract the data.
- Carry out an exploratory analysis of the data with a visual view of the proposed objectives and scope. From the results, limit the scope of the project and define the success metrics.
- Identify the features with the experts of the vertical.
- Present a general outline of features, architectures, models and performance metrics.
- Process and transform the data, extract the features, divide the data into training, validate and test.
- Train the model and adjust the hyperparameters.
- Select the model, obtain performance, analyze and present the results.
- Start the industrialization process.
Solution: AI in Oil and Gas
Reading and transformation of seismic volume.
A seismic volume is the response of the subsoil to an applied energy. The waves generated by this energy, which can be released through vibrations or explosions, propagate through the subsoil.
The reflected waves are captured by geophones or hydrophones on the surface and after a process of ordering, filtering and correction of traces, the seismic volume is obtained.
It is very common for traces to be stored in SEGY data format, so for their processing as a 3D volume each trace must be read and stacked to form the volume. Then corrections and normalizations should be made to aid in the ML model training process.
Training data generation
The seismic volume is made up of inline and crossline images throughout the depth of the volume. In order for neural network architecture to learn, it is necessary to show you a set of examples of the original volume and its correct segmentation carried out by an expert.
In our methodology, only the previous segmentation of 1-3% of the volume is necessary. In this way thousands of pieces of the volume are chosen to train the model.
Construction of the deep RNA model or architecture
These types of models have millions of parameters that must be determined, iteratively, in the training process.
For this reason, we make use of GPUs and TPUs (Tensor Processing Units) to speed up the process.
Performance measurement and results analysis
The best results obtained, with different seismic volumes, give an accuracy close to 98%.
This result is very close to the best performances reported in the most recent literature.
In the prediction process, the geologist and geophysicist should only use the previously trained model to perform the automatic picking of the seismic volume.
This allows you to save time, make more precise interpretations, which in the end translates into economic benefits for the company.